Behavior Observation with Motion Trajectory Description

Effective observation and understanding of behaviors is of critical importance for visual surveillance and the related applications. However, the raw data from motion trajectories have been mostly used for behavior representation in the previous work, which is inflexible to use. To overcome this difficulty, we propose a flexible and generic motion trajectory descriptor to serve as an effective behavior description mechanism. The descriptor is not only advantageous in offering a generalized and reusable behavior description, but also can boost the behavior understanding through adaptive behavior recognition, anomaly detection and behavior prediction. It is foreseeable that the systematic behavior descriptor can promise wide applications in practice. This project also aims at developing a novel solution for modeling and understanding human behaviors via proposing a theoretically complete motion trajectory signature descriptor. A mode-based method for behavior description will be studied and Gaussian mixture model will be explored to build an abstract description for semantic simple behaviors. The high-level generative model will be adopted to describe the semantic rich multi-episode human behaviors. In addition, we will study the invariants of the behavior description. The proposed implementation system will be useful for many applications of advanced surveillance involving behavior recognition such as robot learning by demonstration, human activity monitoring and security control.